Re: [PATCH] c++: Implement __is_{nothrow_,}convertible [PR106784]

[Marek Polacek <polacek@redhat.com>]

On Fri, Sep 23, 2022 at 11:54:53AM -0400, Jason Merrill wrote:
> On 9/23/22 10:34, Marek Polacek wrote:
> > On Thu, Sep 22, 2022 at 06:14:44PM -0400, Jason Merrill wrote:
> > > On 9/22/22 09:39, Marek Polacek wrote:
> > > > To improve compile times, the C++ library could use compiler built-ins
> > > > rather than implementing std::is_convertible (and _nothrow) as class
> > > > templates.  This patch adds the built-ins.  We already have
> > > > __is_constructible and __is_assignable, and the nothrow forms of those.
> > > > 
> > > > Microsoft (and clang, for compatibility) also provide an alias called
> > > > __is_convertible_to.  I did not add it, but it would be trivial to do
> > > > so.
> > > > 
> > > > I noticed that our __is_assignable doesn't implement the "Access checks
> > > > are performed as if from a context unrelated to either type" requirement,
> > > > therefore std::is_assignable / __is_assignable give two different results
> > > > here:
> > > > 
> > > >     class S {
> > > >       operator int();
> > > >       friend void g(); // #1
> > > >     };
> > > > 
> > > >     void
> > > >     g ()
> > > >     {
> > > >       // #1 doesn't matter
> > > >       static_assert(std::is_assignable<int&, S>::value, "");
> > > >       static_assert(__is_assignable(int&, S), "");
> > > >     }
> > > > 
> > > > This is not a problem if __is_assignable is not meant to be used by
> > > > the users.
> > > 
> > > That's fine, it's not.
> > Okay then.  libstdc++ needs to make sure then that it's handled right.
> > 
> > > > This patch doesn't make libstdc++ use the new built-ins, but I had to
> > > > rename a class otherwise its name would clash with the new built-in.
> > > 
> > > Sigh, that's going to be a hassle when comparing compiler versions on
> > > preprocessed code.
> > 
> > Yeah, I guess :/.  Kind of like __integer_pack / __make_integer_seq.
> > 
> > > > --- a/gcc/cp/constraint.cc
> > > > +++ b/gcc/cp/constraint.cc
> > > > @@ -3697,6 +3697,12 @@ diagnose_trait_expr (tree expr, tree args)
> > > >        case CPTK_HAS_UNIQUE_OBJ_REPRESENTATIONS:
> > > >          inform (loc, "  %qT does not have unique object representations", t1);
> > > >          break;
> > > > +    case CPTK_IS_CONVERTIBLE:
> > > > +      inform (loc, "  %qT is not convertible from %qE", t2, t1);
> > > > +      break;
> > > > +    case CPTK_IS_NOTHROW_CONVERTIBLE:
> > > > +	inform (loc, "  %qT is not %<nothrow%> convertible from %qE", t2, t1);
> > > 
> > > It's odd that the existing diagnostics quote "nothrow", which is not a
> > > keyword.  I wonder why these library traits didn't use "noexcept"?
> > 
> > Eh, yeah, only "throw" is.  The quotes were deliberately added in
> > <https://gcc.gnu.org/pipermail/gcc-patches/2019-May/522333.html>.  Should
> > I prepare a separate patch to use "%<noexcept%>" rather than "%<nothrow%>"?
> > OTOH, the traits have "nothrow" in their names, so maybe just go back to
> > "nothrow"?
> 
> The latter, I think.  Or possibly "no-throw".  I guess -Wformat-diag wants
> "nothrow" quoted because of the attribute of that name.

OK, let me see.
 
> > > > --- a/gcc/cp/method.cc
> > > > +++ b/gcc/cp/method.cc
> > > > @@ -2236,6 +2236,37 @@ ref_xes_from_temporary (tree to, tree from, bool direct_init_p)
> > > >      return ref_conv_binds_directly (to, val, direct_init_p).is_false ();
> > > >    }
> > > > +/* Return true if FROM can be converted to TO using implicit conversions,
> > > > +   or both FROM and TO are possibly cv-qualified void.  NB: This doesn't
> > > > +   implement the "Access checks are performed as if from a context unrelated
> > > > +   to either type" restriction.  */
> > > > +
> > > > +bool
> > > > +is_convertible (tree from, tree to)
> > > 
> > > You didn't want to add conversion to is*_xible?
> > 
> > No, it didn't look like a good fit.  It does things we don't need, and
> > also has if VOID_TYPE_P -> return error_mark_node; which would be wrong
> > for __is_convertible.
> > 
> > I realized I'm not testing passing an incomplete type to the built-in,
> > but since that is UB, I reckon we don't need to test it (we issue
> > "error: invalid use of incomplete type").
> 
> But your patch does test that, in the existing call to check_trait_type from
> finish_trait_expr?

Yes, it eventually checks complete_type_or_else.

> The patch is OK.

Thanks,

Marek